ARBA deployment check

parent project ← Pathway satisfiability

A runnable prototype that applies the module satisfiability engine to automated
biological-process calls from UniProt's ARBA rules (GO_REF:0000117).

The idea

An ARBA process call made from one sequence bundles two inferences:

1. sequence → activity. Does the protein catalyse the reaction? This is local
   to the sequence (family, FunFam, catalytic residues).
2. activity → process ("deployment"). Is that activity used in the process
   in this organism? This depends on the rest of the genome: whether the other
   steps exist, whether a paralog does the job, whether the enzyme's role lies
   elsewhere.

GO does not encode the second inference. For example, GO:0004617
phosphoglycerate dehydrogenase activity has no part_of link to GO:0006564
L-serine biosynthetic process; its only non-subclass relations are its
substrates and products. Curated modules do encode it, so the check makes it
explicit:

Protein P is involved_in process T in organism O if (a) P's own
molecular-function annotations fill a step of a module whose process concept
is T or a descendant of T, and (b) that module is satisfiable in O. A step
holds if any O protein has a GOA annotation to the step's activity (QuickGO,
any evidence, is_a descendants included).

Modules with context.taxa apply only to organisms inside those taxa.

exact_concept in results.tsv records whether the module's process concept is
the row's term itself (True) rather than a descendant of it.

Results

Current output is in RESULTS.md; per-row detail is in
results.tsv. The figures below are from the run of 2026-10-17
(4,058 ARBA process rows in reviewed genes; 1,617 scored), and RESULTS.md
emits the same table on every run.

Share of rows the gene review kept (ACCEPT or KEEP_AS_NON_CORE) or rejected
(REMOVE, MARK_AS_OVER_ANNOTATED or TOO_GENERAL), by verdict:

What the numbers say:

• An exact entailment is a strong positive signal. When the protein fills a
  step of a satisfiable module whose process is exactly the ARBA term, reviewers
  kept 93% of the rows and removed none outright (8 of 220 were marked
  over-annotated). These rows could be generated from the activity call plus the
  module, which states the grounds for the call.
• Inherited entailment is true but often uninformative. A protein that fills
  a step of a satisfiable module is involved in every ancestor of that module's
  process. Reviewers still reject 19% of these rows, mostly as over-annotation
  of vague terms (system development, response to stress, animal organ
  development). This check tests whether a call is grounded, not whether the term
  is specific enough; a specificity filter has to be applied separately.
• NOT_ENTAILED means no evidence, not evidence against. Most such rows are
  kept by reviewers. Typical causes: no module covers the protein's actual role
  (chaperones for response to heat, protein refolding), or the protein's GOA
  activity terms do not reach the module's step term.
• MODULE_INCOMPLETE is mostly a diagnostic for modules and the oracle. Two
  recurring causes:
• Modules without context.taxa that are in practice lineage-specific (e.g.
  bacterial_recbcd_end_resection scored against human RAD51 paralogs). Adding
  taxa to these modules would remove the noise.
• Steps whose activity term is not what GOA uses for the organism's own enzyme
  (e.g. c_elegans_daf2_insulin_signaling is unsatisfiable in C. elegans at
  the AGE-1 and PDK-1 steps). These point at a module term choice or a GOA
  activity gap worth checking.

The case that motivated this

P. putida serA (Q88CM5) and M. extorquens serA (C5APD6) receive
GO:0006564 L-serine biosynthetic process only from ARBA (rule
ARBA00093701, a single CATH FunFam condition); no other pipeline reaches it.
Both are ENTAILED: each fills the phosphoglycerate dehydrogenase step of
phosphorylated_serine_biosynthesis, and the SerC and SerB steps are filled by
other proteins in each proteome. The ARBA call is therefore backed by the module
rather than being asserted without stated grounds.

Caveats

• The oracle uses GOA activity annotations, including electronic ones (ARBA,
  EC, Rhea, InterPro2GO). That is deliberate (step 1 is the activity
  inference), but a wrong activity call propagates into the verdict.
• Steps without a GO activity term cannot be tested and are assumed present.
• Presence is not flux: a satisfiable module is permitted, not shown to run.
• Taxon scope comes from each module's context.taxa. A module that declares
  no taxa is scored for every organism, so a lineage-specific module without
  taxa can be matched across kingdoms; this affects ENTAILED and NOT_ENTAILED as
  well as MODULE_INCOMPLETE. Before the dicty_* and pqq_biosynthesis modules
  were given taxa, zebrafish shha was ENTAILED for regulation of developmental
  process solely by the Dictyostelium counting-factor module, and human
  CYP21A2 was scored against bacterial PQQ biosynthesis. A module that declares taxa only as free text (no
  NCBITaxon id) is excluded, and RESULTS.md lists those modules; grounding their
  taxa would bring them back.
• The organism-level oracle is stricter than the gene-level test: it asks
  QuickGO for is_a descendants of the step activity with an exact taxon match,
  while the protein's own activity is matched through is_a and part_of
  ancestors. A step can therefore read as absent when its enzyme is annotated
  only at strain level or through part_of.
• Competing modules are not yet weighed. other_modules lists modules outside
  T that the protein also fills (e.g. a dehydrogenase that also fits vitamin B6
  biosynthesis), for manual review.

Reproduce

uv run python projects/PATHWAY_SATISFIABILITY/arba_deployment/deploy_check.py

QuickGO and UniProt taxonomy responses are cached under .cache/arba_deployment/
(git-ignored); a first run takes about 20 minutes and makes roughly 11,000
QuickGO requests.